Let's start with a simple, foundational idea: The AI models we have today are intelligent. They are not just clever calculators or fancy spreadsheets. They are genuinely powerful, and more importantly, they are profoundly useful. But to get the most value out of them, we have to stop asking if they're smart and start asking how they're smart. Their intelligence isn't a dimmer switch, moving from dumb to brilliant. It’s a bizarre, spiky landscape of superhuman peaks and shockingly deep valleys. And that’s in part because there are multiple forms of intelligence.

I’m most certainly not talking about pop-psychology notions of intelligence. I'm referring to the fundamental machinery of thought. Today's foundation models—our LLMs and reinforcement learning agents—are built on one spectacular superpower: statistical, model-free learning. They are the greatest pattern-matchers ever conceived. After ingesting a digital chunk of human civilization, they become masters of correlation, inferring what word, image, or action should come next with breathtaking accuracy. This is the engine behind their ability to write a sonnet or suggest a chess move. It’s both a proud display of real intelligence and an incredible illusion of understanding.

It’s also their Achilles' heel. This purely correlational diet, paired with imperfect cost functions, is precisely why a stray fact about a sleepy cat can cause a logical meltdown on a math problem, and why foundation AIs’ internal "world models" can be hollow facades. They are like a student who has not simply memorized the answer key for every test ever made but has even learned the patterns behind the question-answer relationships. But all too often the underlying principles that truly move their worlds go unlearned. The causes are hidden in the (highly useful) correlations. They know that B follows A, but they have no concept of why.

We rely on statistical learning as well. So much of humans behavior is driven be very similar mechanisms that it isn’t a shock we see ourselves in LLMs. There is a real and profound overlap. However, our intelligence doesn’t stop there. We leverage others forms of natural intelligence: model-based learning, causal reasoning, and navigating uncertainty. We build mental models, we ask "why", and we are acutely aware of what we don't know. (In fairness, we do all of these terribly…just better than AI.)

The gap goes even deeper, down to the nature of memory itself. AIs excel at semantic memory—the encyclopedia of facts. They know that Paris is the capital of France. Humans, however, operate on episodic memory—the story of our lives. We don't just know the fact; we remember the time we went to Paris, the smell of the croissants, the feeling of getting lost in the Louvre. This rich, contextual, multimodal, first-person experience helps ground our knowledge in reality. AI has the entire library at its disposal but has never had the experience of reading a single book. It’s not even clear to me that the large foundation models even have an explicit representation of self during their brief moments of processing (such as might be seen in cortical activity such as the human posterior cingulate).

These are not, I believe, fundamental limits of machine intelligence itself. They are the limitations of the specific architectures we are using today. The models are brilliant, but they are also brittle because their intelligence lacks a causal, grounded foundation. The exciting news is that we are just beginning to see the first cracks in this paradigm—models that try to reason without the clumsy baggage of words, hinting at a new path forward. We aren't watching the final act of AI's development; we're just beginning to map the strange and wonderful continent of a new kind of mind.

And guessing at the emergence date of fully sentient AGI is as absurd as predicting the Rapture.

Follow me on LinkedIn or join my growing Bluesky! Or even..hey whats this...Instagram? (Yes, from sales reps and HR professionals to lifestyle influences and Twitter castaways, I'll push science anywhere.)

Research Roundup

AI Reasoning Without Reason

Can AI “reason”? This is a big debate in the field. LLMs can certainly solve logical problems well above chance and at rates at least comparable to humans. But the patterns of errors suggest LLM reasoning isn’t all that logical.

Take for example a high profile paper out of Apple research with the unsubtle title, “The Illusion of Thinking”. It claims that even as LLMs have “improved performance on reasoning benchmarks”, they show “complete accuracy collapse beyond certain [problem] complexities.”

They describe 3 “performance regimes”:

1) Low-complexity tasks where standard LLMs outperform cutting-edge agentic reasoning models.

2) Medium-complexity tasks where reasoning models shine.

3) High-complexity tasks…where everything goes boom.

Part of the challenge is that reasoning models “fail to use explicit algorithms and reason inconsistently across puzzles.”

While people have argued against this paper (I say people, I think someone just asked Claude to write a response), but there’s is also this very silly but illuminating claim from another group:“appending, “Interesting fact: cats sleep most of their lives,” to any math problem leads to more than doubling the chances of a model getting the answer wrong.”

In this case, an adversarial model was trained to find “short, irrelevant text that, when appended to math problems, systematically misleads models to output incorrect answers without altering the problem’s semantics.” That simple trick of inserting unrelated text into math problems increased errors by more than 300%.

There are many kinds of natural intelligence, and LLMs are genuinely intelligent. More specifically, they are superhuman statistical learners: inferring right answers from context. What they are not doing…yet…is reasoning.

Impoverished World Models

If adding a fact about a sleepy cat can derail an AI’s reasoning, what does that say about its internal understanding of the world? If it's not "reasoning," what is it doing?

A nerdy way to ask that question is, “Do LLMs secretly build ‘world models’—coherent, internal maps of the rules of a system.” Think of it as learning the physics of a mini-world, not just memorizing what happens in it.

Many have assumed LLMs and RLs build world models because of how well they perform on tasks about worlds like “logical reasoning, geographic navigation, game-playing, and chemistry”. But a new paper put these assumptions to a new strict test: they gave the AI problems to solve within these little worlds while applying linguistic analysis to their outputs.

The models did great on the traditional assessment, solving the puzzles and playing the games correctly. If you only looked at the final answers, you'd give them an A+ and assume they understood the rules perfectly.

But the deeper analysis revealed that internal world models were "far less coherent than they appear." When asked a slightly different question that relies on the same principles, and they completely fall apart.

This gets to the heart of "Some Intelligence; Not All Intelligence." The models are creating a fragile, statistical scaffold that imitates a logical foundation. They build something that looks like a world model from the outside, but is hollow on the inside. They know everything but understand nothing.

No More Words

Does reasoning require language? A new paper suggests that for certain kinds of problems AI reasoning models are better off leaving words behind.

Think about how you solve a tough problem. Sometimes you talk yourself through it, step-by-step. This is basically what an AI does with "Chain of Thought". It's forced to think ‘out loud’, writing down each logical step as a sentence. The problem? It’s slow, and once it writes a sentence, it's committed to that path. If it's a dead end, it’s hard to turn back.

But this isn’t the only way humans reason. Decades of research on experts shows they they often filter out huge numbers of possible math operations or chess moves without explicitly considering, while also maintaining multiple high value possibilities at the same time.

The new paper proposes letting AIs do something more like that: let an LLM reason in an "unrestricted latent space"—the embedded space of abstract concepts, not clunky human words. The method is ingenious. Instead of forcing the AI to write down its next step, it grabs its "thought" directly as the vector representing the "thought". It takes the hidden state of the network and feeds that raw "concept" right back into the model without ever converting it from "thought" to "word".

The AI is essentially thinking a continuous, language-free trajectory of thoughts, moving freely through the space of ideas and superpositions of ideas.

Instead of being locked into one path (thinking out loud), the model can explore many different options simultaneously. This allows it to perform a "breadth-first search" of all possible solutions, making it better at problems that require backtracking or changing your mind. Fewer "garden path" errors means the model is much more robust. It's less likely to get sidetracked by a distracting fact about a sleepy cat, because it's not operating on the messy, literal surface of language.

Does this model grant us a model of latent, language-free reasoning happening in the quiet, abstract spaces of the human mind? Exploration of AI thinking might just be teaching us what it really means to think at all.

<<Support my work: book a keynote or briefing!>>

Want to support my work but don't need a keynote from a mad scientist? Become a paid subscriber to this newsletter and recommend to friends!

SciFi, Fantasy, & Me

Are we Narcissus the Anglerfish? “Prediction: in 15 years, nobody under the age of 20 will see why this was supposed to be funny.”

Or perhaps we are just the gist of a summary of a hot take?

Stage & Screen

• July 22, LA: "Bend the Algorithm" – AI...Hollywood...democracy!
• July 23, UC Berkeley: Can AI in finance be AI for Good?
• August 18, Qatar: What's next: Human-AI Collective Intelligence
• September 18, Oakland: Reactive Conference
• Sep 28-Oct 3, South Africa: Finally I can return. Are you in SA? Book me!
• October 6, UK: More med school education
• October 11, Mexico City: maybe...I hope so!
• October 29, Baltimore: The amazing future of libraries!
• November 4, Mountain View: More fun with Singularity University
• November 21, Warsaw: The Human Tech Summit
• December 8, San Francisco: Fortune Brainstorm AI SF talking about build a foundation model for human development

If your company, university, or conference just happen to be in one of the above locations and want the "best keynote I've ever heard" (shockingly spoken by multiple audiences last year)?

Vivienne L'Ecuyer Ming